I've noticed a difference between the output of the GRN layer using the CPU plugin and the Myriad plugin. The CPU plugin behaves as I would expect, by dividing the channel values by the L2 norm of the channel vector, but the Myriad plugin appears to divide by the absolute value of each channel value, creating a vector of +1 and -1 values. Here are two notebooks showing my test results:
I've tested on a few devices so I believe the problem is in the plugin.
Python Version: 3.6.7
OpenVINO version: 2018.5.455
Operating System: Ubuntu 18.04
Have other's noticed this or can you reproduce it?
I recently upgraded to the newest version, 2019 R1 133 and I have the same issue with the Myriad plugin and Global Response Normalization. The above gists produce the same results as the previous version.
Can the internal devs at intel reproduce the issue?